1 / 64

CSCI 4717/5717 Computer Architecture

CSCI 4717/5717 Computer Architecture. Topic: Symmetric Multiprocessors & Clusters Reading: Stallings, Sections 18.1 through 18.4. Classifications of Parallel Processing.

Download Presentation

CSCI 4717/5717 Computer Architecture

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CSCI 4717/5717 Computer Architecture Topic: Symmetric Multiprocessors & Clusters Reading: Stallings, Sections 18.1 through 18.4

  2. Classifications of Parallel Processing M. Flynn classified types of parallel processing in 1972 ("Some Computer Organizations and Their Effectiveness", IEEE Transactions on Computers) Types of Parallel Processor Systems (Figure 18.2) • Single instruction, single data stream • Single instruction, multiple data stream • Multiple instruction, single data stream • Multiple instruction, multiple data stream

  3. Classifications of Parallel Processing (continued) • Single Instruction, Single Data Stream (SISD) – Single processor operates on a single instruction stream from a single memory (Uniprocessor) • Single Instruction, Multiple Data Stream (SIMD) – Lockstep operation of multiple processors on single instruction memory with one data memory per processing element. (Vector/array processing)

  4. Classifications of Parallel Processing (continued) • Multiple Instruction, Single Data Stream (MISD) – Multiple processors execute different sequences of instructions on a single data set. Not commercially implemented • Multiple Instruction, Multiple Data Stream (MIMD) – A set of processors simultaneously execute different instructions on different data sets.

  5. Classifications of Parallel Processing (continued)

  6. Multiple Instruction, Multiple Data Stream • Processors are general purpose • Each processor should be able to complete process by themselves • Communications methods • Through shared memory ("Tightly Coupled") • Symmetric multiprocessor (SMP) – memory access times are consistent for all processors • Nonuniform Memory Access (NUMA) – memory access times may differ • Cluster – Either through fixed connections or a network ("Loosely Coupled")

  7. Symmetric Multiprocessors (SMP) • Originally for science, industry, and business where custom multithreaded software could be written • Stand alone computer with the following traits • 2 or more processors of comparable capacity • Processors share same memory and I/O • Processors are connected by a bus or other internal connection • Memory access time is approximately the same for each processor

  8. Symmetric Multiprocessors (continued) • All processors share access to I/O through either: • same channels • different channels providing paths to same devices • All processors can perform the same functions (hence symmetric) • System controlled by integrated operating systemproviding interaction between processors • Interaction at job, task, file and data element levels • Performance improvement of SMP will not be a multiple of the number of processors – choked by RAM access and cache coherence

  9. Symmetric Multiprocessor (continued) • Consumer apps (e.g., gaming) not usually written for SMP • Majority of consumer machines are uniprocessor • Running SMP app on uniprocessor system results in poor performance • For maximum performance in consumer apps, need two development paths: SMP and uniprocessor • SMP still improves uniprocessor applications • One processor can run app while other processor takes care of interrupts and O/S overhead

  10. SMP Advantages Advantages only realized if O/S can provide parallelism • Performance, but only if some work can be done in parallel • Availability/reliability – Since all processors can perform the same functions, failure of a single processor does not halt the system • Incremental growth – User can enhance performance by adding additional processors • Scaling – Vendors can offer range of products based on number of processors • Transparent to user – User only sees improvement in performance

  11. Organization of Tightly Coupled Multiprocessor • Individual processors are self-contained, i.e., they have their own control unit, ALU, registers, one or more levels of cache, and private main memory • Access to shared memory and I/O devices through some interconnection network • Processors communicate through memory in common data area

  12. Organization of Tightly Coupled Multiprocessor (continued) • Memory is often organized to provide simultaneous access to separate blocks of memory • Bus • Time-shared or common bus • Multiport memory • Central controller (arbitrator)

  13. Organization of Tightly Coupled Multiprocessor (continued)

  14. Time Shared Bus • Simple extension of single processor bus • Can be likened to relationship between DMA and single processor • Following features provided • Addressing – distinguish modules on bus • Arbitration • Any module can be temporary master • Must have an arbitration scheme • Time sharing – if one module has the bus, others must wait and may have to suspend • Now have multiple processors as well as multiple I/O modules on a single bus

  15. Time Shared Bus (continued)

  16. Time Shared Bus (continued) • Advantages • Simplicity – not only is it easy to understand, but its use has been well-tested with DMA • Flexibility – adding processor involves simple addition of processor to bus • Reliability – As long as arbitration does not involve single controller, then there is no single point of failure

  17. Time Shared Bus (continued) • Disadvantages • Waiting for bus creates bottleneck • Can be helped with individual caches • Usually L1 and L2 • Cache coherence policy must be used (usually hardware)

  18. Multiport Memory • Direct independent access of memory modules by each processor and I/O module • Memory handles "arbitration" – Logic internal to memory required to resolve conflicts • Little or no modification to processors used to single processor applications or I/O modules required

  19. Multiport Memory (continued)

  20. Multiport Memory (continued) • Advantages • Removing bus access bottleneck • Dedicate portions of memory to only one processor • Better security • Better recovery from faults • Disadvantages • Complex memory logic • More PCB wiring – basically, there's a unique bus for each processor • Write through policy should be used for caches • Central point of failure

  21. Central Control Unit • A separate processor called a central control unit handles the integration of the processors • Functions of central control unit • Funnels separate data streams between independent modules • Can buffer requests • Performs arbitration and timing • Pass status and control • Perform cache update alerting

  22. Central Control Unit (continued) • Uses same control, addressing, and data interfaces as a typical processor, therefore, interfaces to modules remain the same • Disadvantages • Very complex control unit • Control unit is possible bottleneck • Central point of failure

  23. SMP Integrated Operating System • O/S must provide SMP support or only one processor will be used, others ignored • O/S for SMP is NOT like clusters/loosely coupled where communication usually is at file level • Can be a high degree of interaction between processes

  24. SMP Operating System (continued) • To user, it appears as if there is a single O/S, i.e., single processor multiprogramming system • O/S schedules processes or threads across all processors • User should be able to create multithreaded processes without needing to know whether one processor or more will be used

  25. SMP Operating System Design Issues • Simultaneous concurrent processes • O/S routines should be reentrant • O/S tables and other management structures must be expanded to handle multiple processes and processors • Scheduling • More than just order now, also which processor gets a process • Any processor should be capable of scheduling too

  26. SMP Operating System Design Issues (continued) • Synchronization – scheduling of resources now more than just for processes but also for processors • Memory management • Shared page replacement strategy • Must understand and take advantage of memory hardware • Reliability and fault tolerance – Must be able to handle the loss of a processor without taking down other processors.

  27. Cache Coherence • One or two levels of cache typically associated with each processor – this is essential for performance • Problem • Multiple copies of same data in different caches • Can result in an inconsistent view of memory

  28. Write Policy Review • Write back policy • Write goes only to cache • Main memory updated only when cache block is replaced • Can lead to inconsistency • Write through policy • All writes made to cache and main memory • Inconsistencies can occur unless all caches monitor memory traffic

  29. Software Solutions • Compiler and operating system deal with problem • Overhead transferred to compile time • Design complexity transferred from hardware to software • Software tends to make conservative decisions leading to inefficient cache utilization

  30. Software Solutions (continued) • Marked shared variables as non-cacheable – Too conservative • Instructions added to enable/disable caching for variables. Then compiler can analyze code to determine safe periods for caching shared variables

  31. Hardware Solution • A.K.A cache coherence protocols • Dynamic recognition of potential problems at run time • Because it only deals w/problem when it occurs, more efficient use of cache • Transparent to programmer and compiler • Methods • Directory protocols • Snoopy protocols

  32. Directory Protocols – Central control • Central memory controller maintains directory of: • where blocks are held • in which caches they are held • what state the data is in • Appropriate transfers are performed by controller

  33. Directory Protocols – Write Process • Requests to write to a line are made to controller • Using directory, controller tells all other processors with copy of same data to invalidate • Write is granted to requesting processor and that processor has exclusive rights to that data • Request to read from another processor forces controller to issue command to processor with exclusive rights to update (write back) main memory.

  34. Directory Protocols (continued) • Problems • Creates central bottleneck • Communication overhead • Effective in large scale systems with complex interconnection schemes

  35. Snoopy Protocols – Distributed control • Distribute cache coherence responsibility among cache controllers • Cache recognizes that a line is shared • Updates announced to other caches • Suited to bus based multiprocessor • Problem – possible to increase bus traffic to point of canceling out benefits • Two types of implementations: • Write Invalidate • Write Update

  36. Write Invalidate (a.k.a. MESI) • Multiple readers, one writer • When a write is required, command is issued and all other caches of the line are invalidated • Writing processor then has exclusive (cheap) access until line required by another processor • A state is associated with every line • Modified • Exclusive • Shared • Invalid

  37. Write Update (a.k.a. write broadcast) • Multiple readers and writers • Updated word is distributed to all other processors

  38. Snoopy Protocols – Implementations • Performance of these two implementations depends on number of caches and pattern of read/writes • Some systems use adaptive protocols to use both methods • Write invalidate most common – Used in Pentium 4 and PowerPC systems

  39. MESI Protocol • Each line of a cache has associated with it two bits – four states • Modified – line in this cache is modified and only valid in this cache • Exclusive – line in this cache is same as that in memory (unmodified) and not present in any other cache • Shared – line in this cache is same as that in memory (unmodified) and may also be present in another cache • Invalid – line in this cache contains bad data • Write throughs from an L1 cache to an L2 cache makes it visible to the MESI protocol

  40. MESI Protocol (continued)

  41. MESI – State Transition Diagram

  42. Clusters • Defined • a group of interconnected, whole computers • working together as a unified computing resource • can create the illusion of being one machine • Alternative to Symmetric Multiprocessing (SMP) • High performance • High availability • Server applications • Each computer called a node

  43. Cluster Computer Architecture Figure 18.11 from Stallings, Computer Organization & Architecture

  44. Cluster Benefits • Absolute scalability – Almost limitless in terms of adding independent multiprocessing machines • Incremental scalability – Can start out small and build as user acquires new machines

  45. Cluster Benefits (continued) • High availability • Loss of one node only causes small decrement in performance • Software (middleware) handles fault tolerance automatically • Superior price/performance • By using easily affordable building blocks, gets better performance at a lower price than a single large computer • Expanding design doesn't depend on PCB redesign

  46. Cluster Configurations • High-speed message link options/configurations • Dedicated LAN with at least one having connection to remote client • Shared LAN with other non-cluster machines • Simplest way to classify clusters is based on whether computers share disk(s) • No shared disk – each machine has a local disk • Shared disk in addition to local disk – should use disk mirroring or RAID

  47. Cluster Configurations – Standby Server with no Shared Disk

  48. Cluster Configurations – Shared Disk

  49. Cluster Configurations (continued) • Secondary server – cluster functional classification • Passive Standby • Second computer will take over in the event of a failure on the part of the first • First computer sends "heartbeat“ • Heartbeat stops, secondary takes over • Data must be shared or disks must be shared in order for secondary to take over database stuff too • Active Standby – Second computer participates in processing

  50. Active Standby Configurations • Separate Server • No shared disk • High performance and availability • Scheduling software is needed to assign client requests to servers to balance the load • If a computer fails in middle of application, another can take over • To do this, must have some method of copying data between at least neighboring computers

More Related